Narrowing conversion differences

With both g++ and clang it's ok to compile

std::vector<int> xs{1, 2, 3u};

but it's not ok to compile

std::vector<int> xs{1, 2, 3.0};

What is the rationale behind this apparently bizarre choice?

If we consider the type of the last value both unsigned and double cannot be converted safely to an int, and if we consider instead the specific literal values both 3u and 3.0 can instead be converted safely.

Why the distinction then?